[next-20150119] regression (mm)?

On Fri, Jan 23, 2015 at 02:42:17PM -0800, Tyler Baker wrote:
> Hi Kirill,
> 
> On 23 January 2015 at 12:22, Kirill A. Shutemov <kirill@shutemov.name> wrote:
> > On Fri, Jan 23, 2015 at 12:37:06PM -0600, Nishanth Menon wrote:
> >> On 09:39-20150123, Tyler Baker wrote:
> >> > Hi,
> >> >
> >> > On 23 January 2015 at 09:27, Nishanth Menon <nm@ti.com> wrote:
> >> > > On 16:05-20150120, Kirill A. Shutemov wrote:
> >> > > [..]
> >> > >> Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> >> > >> Reported-by: Nishanth Menon <nm@ti.com>
> >> > > Just to close on this thread:
> >> > > https://github.com/nmenon/kernel-test-logs/tree/next-20150123 looks good
> >> > > and back to old status. Thank you folks for all the help.
> >> >
> >> > I just reviewed the boot logs for next-20150123 and there still seems
> >> > to be a related issue. I've been boot testing
> >> > multi_v7_defconfig+CONFIG_ARM_LPAE=y kernel configurations which still
> >> > seem broken.
> >> >
> >> > For example here are two boots with exynos5250-arndale, one with
> >> > multi_v7_defconfig+CONFIG_ARM_LPAE=y [1] and the other with
> >> > multi_v7_defconfig[2]. You can see the kernel configurations with
> >> > CONFIG_ARM_LPAE=y show the splat:
> >> >
> >> > [   14.605950] ------------[ cut here ]------------
> >> > [   14.609163] WARNING: CPU: 1 PID: 63 at ../mm/mmap.c:2858
> >> > exit_mmap+0x1b8/0x224()
> >> > [   14.616548] Modules linked in:
> >> > [   14.619553] CPU: 1 PID: 63 Comm: init Not tainted 3.19.0-rc5-next-20150123 #1
> >> > [   14.626713] Hardware name: SAMSUNG EXYNOS (Flattened Device Tree)
> >> > [   14.632830] [] (unwind_backtrace) from [] (show_stack+0x10/0x14)
> >> > [   14.640473] [] (show_stack) from [] (dump_stack+0x78/0x94)
> >> > [   14.647678] [] (dump_stack) from [] (warn_slowpath_common+0x74/0xb0)
> >> > [   14.655744] [] (warn_slowpath_common) from [] (warn_slowpath_null+0x1c/0x24)
> >> > [   14.664510] [] (warn_slowpath_null) from [] (exit_mmap+0x1b8/0x224)
> >> > [   14.672497] [] (exit_mmap) from [] (mmput+0x40/0xf8)
> >> > [   14.679180] [] (mmput) from [] (flush_old_exec+0x328/0x604)
> >> > [   14.686471] [] (flush_old_exec) from [] (load_elf_binary+0x26c/0x11f4)
> >> > [   14.694715] [] (load_elf_binary) from [] (search_binary_handler+0x98/0x244)
> >> > [   14.703395] [] (search_binary_handler) from []
> >> > (do_execveat_common+0x4dc/0x5bc)
> >> > [   14.712421] [] (do_execveat_common) from [] (do_execve+0x28/0x30)
> >> > [   14.720235] [] (do_execve) from [] (ret_fast_syscall+0x0/0x34)
> >> > [   14.727782] ---[ end trace 5e3ca48b454c7e0a ]---
> >> > [   14.733758] ------------[ cut here ]------------
> >> >
> >> > Has anyone else tested with CONFIG_ARM_LPAE=y that can confirm my findings?
> >> Uggh... I missed since i was looking at non LPAE omap2plus_defconfig.
> >>
> >> Dual A15 OMAP5432 with multi_v7_defconfig + CONFIG_ARM_LPAE=y
> >> https://github.com/nmenon/kernel-test-logs/blob/next-20150123/multi_lpae_defconfig/omap5-evm.txt
> >>
> >> Dual A15 DRA7/AM572x with same configuration as above.
> >> https://raw.githubusercontent.com/nmenon/kernel-test-logs/next-20150123/multi_lpae_defconfig/dra7xx-evm.txt
> >> https://github.com/nmenon/kernel-test-logs/blob/next-20150123/multi_lpae_defconfig/am57xx-evm.txt
> >>
> >> Single A15 DRA72 with same configuration as above:
> >> https://raw.githubusercontent.com/nmenon/kernel-test-logs/next-20150123/multi_lpae_defconfig/dra72x-evm.txt
> >>
> >> You are right. the issue re-appears with LPAE on :(
> >> Apologies on missing that.
> >
> > Guys, could you instrument mm_{inc,dec}_nr_pmds() with dump_stack() +
> > printk() of the counter and add printk() on mmap_exit() then run a simple
> > program which triggers the issue?
> 
> For reference, here is the patch I've applied for testing, mostly
> stolen from Felipe's debug patch above in this thread.
> 
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 1fbd0e8..e5b0444 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -1455,11 +1455,17 @@ static inline unsigned long mm_nr_pmds(struct
> mm_struct *mm)
>  static inline void mm_inc_nr_pmds(struct mm_struct *mm)
>  {
>         atomic_long_inc(&mm->nr_pmds);
> +        dump_stack();
> +        printk(KERN_INFO "===> %s nr_pmds %ld\n", __func__,
> +                atomic_long_read(&mm->nr_pmds));
>  }
> 
>  static inline void mm_dec_nr_pmds(struct mm_struct *mm)
>  {
>         atomic_long_dec(&mm->nr_pmds);
> +        dump_stack();
> +        printk(KERN_INFO "===> %s nr_pmds %ld\n", __func__,
> +                atomic_long_read(&mm->nr_pmds));
>  }
>  #endif
> 
> diff --git a/mm/mmap.c b/mm/mmap.c
> index 6a7d36d..a16471f 100644
> --- a/mm/mmap.c
> +++ b/mm/mmap.c
> @@ -2809,6 +2809,7 @@ EXPORT_SYMBOL(vm_brk);
>  /* Release all mmaps. */
>  void exit_mmap(struct mm_struct *mm)
>  {
> +       printk(KERN_INFO "===> %s exit_mmap enter\n", __func__);
>         struct mmu_gather tlb;
>         struct vm_area_struct *vma;
>         unsigned long nr_accounted = 0;
> 
> I applied this patch to the tip of linux-next, configured for
> multi_v7_defconfig and set CONFIG_ARM_LPAE=y. The log for this arndale
> boot can be found here [1]. For good measure, I then rebuilt the
> kernel with CONFIG_ARM_LPAE=n and booted the same platform again. This
> log can be found here [2].
> 
> Happy hunting!

Okay, proof of concept patch is below. It's going to break every other
architecture with FIRST_USER_ADDRESS != 0, but I think it's cleaner way to
go.

The problem is that we check nr_ptes/nr_pmds in exit_mmap() which happens
*before* pgd_free(). And if an arch does pte/pmd allocation in pgd_alloc()
and frees them in pgd_free() we see offset in counters by the time of the
checks.

This scenario happens for all archs with FIRST_USER_ADDRESS != 0 and we
tried to work it around by offsetting expected counter value according to
FIRST_USER_ADDRESS for both nr_pte and nr_pmd in exit_mmap().

But ARM with LPAE also has non-zero USER_PGTABLES_CEILING, but upper
addresses occupied with huge pmd entries, so the trick with offsetting
expected counter value will get really ugly: we will have to apply it
nr_pmds, but not nr_ptes.

The proposal is to move the check to check_mm() which happens *after*
pgd_free(). We would need to adjust pgd_free() on all architectures with
non-zero FIRST_USER_ADDRESS to make accouting properly there. But I think
the end result would be cleaner.

Andrew, any comments?

[next-20150119] regression (mm)?

Commit Message

Comments

Patch